MemForest: An event-tree forest for long-term memory¶
Public-observation parsing diagnostic. Event partitioning, progressive merging and anchor-neighborhood retrieval are implemented; term overlap approximates semantics.
Paper information¶
| Field | Details |
|---|---|
| Paper | arXiv 2609.08273 |
| Institution | Shanghai Jiao Tong University (first author). |
| First public release | 2026-09-08 (arXiv v1) |
| Author code | MemForest. |
| Local adapter or method key | memforest |
| Local reproduction code | src/auto_research/agent_research/latest_20260912.py |
Paper method and evidence¶
Partition history by event into separate trees instead of compressing it into one summary. Merge nodes progressively under capacity pressure. Start retrieval from a semantic anchor and expand through neighboring nodes, retaining temporal structure and links between events.
flowchart LR
H[Historical events] --> P[Partition by event]
P --> F[EventTree forest]
F --> M[Merge progressively under capacity pressure]
Q[Current query] --> A[Find anchor]
M --> A --> N[Expand and retrieve neighbors]Original paper figure¶
Figure 2 shows the event-tree memory architecture. From the original paper; copyright remains with the authors.
Local reproduction¶
The compact CPU EventTree selects trees by token Jaccard overlap and actually compacts them. Three-seed metrics. No author embedding model, LLM summarizer, or long-horizon benchmark was run.